Creating models is a labor of love, but it takes a significant amount of time and compute power to get them just right. If you’re enjoying my models, consider fueling my next project with a coffee on <a target="_blank" rel="ugc" href="https://ko-fi.com/koolchh">Ko-fi</a> ☕. Thank you for keeping this project going!<h1 id="animeboysnabla">∇ AnimeBoysNabla ∇</h1><h1 id="introducing-the-noobai-based-versatile-powerhouse-of-my-anime-boys-model-series.-perfect-for-creators-who-demand-variety-and-precision-in-their-husbando-designs!">Introducing the NoobAI-based, versatile powerhouse of my anime boys model series. Perfect for creators who demand variety and precision in their husbando designs!</h1><h3 id="download-my-comfyui-workflow-here">Download my ComfyUI workflow <a target="_blank" rel="ugc" href="https://huggingface.co/Koolchh/AnimeBoysNabla/resolve/main/AnimeBoysNabla.json">here</a></h3><h2 id="inference-guide">🚀 Inference Guide</h2><ul><li>⚠️ Important: This model uses Zero Terminal SNR with V-prediction. Please ensure you are using the correct settings during inference.<ul><li>ComfyUI Users: Add the <code>ModelSamplingDiscrete</code> node into your workflow. Set <code>sampling</code> to <code>v_prediction</code>, <code>zsnr</code> to <code>true</code>.</li><li>Automatic1111 Users: Place the <code>.yaml</code> config file into the model folder. The <code>.yaml</code> file must have the exact same name as the model file, only with the <code>.yaml</code> extension instead of <code>.safetensors</code>. Set <code>Noise schedule for sampling</code> in settings to <code>Zero Terminal SNR</code>.</li></ul></li><li>Prompting: Always begin your prompt with a score tag (e.g. <code>score_9</code>). You can use any of these styles:<ul><li>Tag soup: <code>score_X, tag1, tag2, tag3, ...</code></li><li>Natural language: <code>score_X, [your description here]</code></li><li>Mixed approach: <code>score_X, [description], tag1, tag2, ...</code></li><li>Tip: If the score tags have too much influence on the style, try lowering the weight (e.g., <code>(score_9:0.5)</code>) or removing them entirely.</li></ul></li><li>Negative Prompt: Choose from one of these two presets depending on your needs:<ol><li>Light: <code>score_1, lowres, artistic error, scan artifacts, jpeg artifacts, multiple views, too many watermarks, negative space, blank page</code></li><li>Heavy: <code>score_1, score_2, score_3, lowres, artistic error, film grain, scan artifacts, jpeg artifacts, chromatic aberration, dithering, halftone, screentones, multiple views, logo, too many watermarks, negative space, blank page</code></li></ol></li><li>VAE: Use the built-in VAE. This model uses <a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf/EQ-SDXL-VAE">KBlueLeaf/EQ-SDXL-VAE</a>.</li><li>CFG Scale: A CFG scale of 3 to 5 is recommended. For finer control, I suggest using <a target="_blank" rel="ugc" href="https://github.com/mcmonkeyprojects/sd-dynamic-thresholding">dynamic thresholding</a>.<ul><li>Pro-tip: I use <code>Half Cosine Up</code> for both modes. Set <code>separate_feature_channels</code> to <code>disable</code>, <code>scaling_startpoint</code> to <code>ZERO</code>, and <code>variability_measure</code> to <code>STD</code>.</li></ul></li><li>Resolution: To get started, try these dimensions:<ul><li>Portrait: 832 × 1216</li><li>Square: 1024 × 1024</li><li>Landscape: 1216 × 832</li><li>Some other supported sizes: 768×1344, 768×1280, 896×1152, 960×1088, 1344×768, 1280×768, 1152×896, 1088×960.</li></ul></li></ul><h2 id="training-details">🧪 Training Details</h2>AnimeBoysNabla was fine-tuned from <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1190596">NoobAI V-Pred 1.0</a> using approximately 950k images. The knowledge cutoff is November 2025.The following tags were used during training to help you steer the results toward your desired style.<h3 id="score-tags">Score tags</h3>Each image is tagged with <code>score_X</code>, where <code>X</code> is a range from 1 to 9.<ul><li><code>score_9</code> represents the highest aesthetic quality based on my personal preferences.</li></ul><h3 id="rating-tags">Rating tags</h3><ul><li><code>rating:general</code>: general</li><li><code>rating:sensitive</code>: sensitive</li><li><code>rating:questionable</code>: questionable</li><li><code>rating:explicit</code>: explicit</li></ul><h3 id="year-tags">Year tags</h3>Use <code>year YYYY</code> (ranging from 2005 to 2025) to target specific era styles.<h3 id="training-configurations">Training configurations</h3><ul><li>Hardware: 4 × Nvidia A100 SXM 80GB</li><li>Optimizer: AdamW 8-bit (Weight Decay: 0.1)</li><li>Gradient Accumulation Steps: 8</li><li>Effective Batch Size: 128 (4 × 8 × 4)</li><li>Learning Rates:<ul><li>U-Net: 2e-5</li><li>Text Encoders: 4e-6</li></ul></li><li>LR Schedule: Cosine with 1% minimal LR and 2,000 warmup steps</li><li>Precision: BF16 Mixed Precision</li></ul><h3 id="changes-from-animeboyszeroxl">🔄 Changes from AnimeBoysZeroXL</h3><ul><li>Base Model: Updated to <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1190596">NoobAI V-Pred 1.0</a>.</li><li>VAE: Switched to <a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf/EQ-SDXL-VAE">KBlueLeaf/EQ-SDXL-VAE</a>.</li><li>Dataset Balancing: Reduced repeats for high-score images.</li><li>Learning Rate: Lowered Text Encoder LR and migrated to a Cosine LR scheduler.</li><li>Optimizer: Transitioned to AdamW 8-bit with 0.1 weight decay.</li><li>Precision: Adopted BF16 mixed-precision training.</li><li>Dropout: Increased full caption dropout to 10%.</li></ul><h2 id="license">License</h2>AnimeBoysNabla is a derivative model of <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1190596">NoobAI V-Pred 1.0</a> by Laxhar Lab. Please read their license before using the model.

Files with SHA256: 2e4f8ec774c90ce84b6ff4e82391c7ce17ad04bf0f9a120e26993c2d0f3ab84a